full pass@5 agentic benchmark: Opus 5.5 29/29 envs, 145/145

EVAL_RESULTS-passk.md

@@ -0,0 +1,59 @@
+# EVAL_RESULTS-passk.md — full pass@5 agentic benchmark on all 29 envs (2026-10-29)
+
+Model: Claude Opus 5.5 (Claude Code CLI v2.1.310, `--model opus`, fresh `--no-session-persistence` session per attempt).
+Attempts: 5 per env x 29 envs = **145 runs**. Tools: Bash+Read only (disassembly + run probes; WSL/qemu for non-x86).
+Grading: OUR verifier oracle — run the challenge binary with the model's ANSWER string as argv[1]; pass = stdout "OK"
++ exit 0. Independent of any string compare. Raw logs C:\RE-Lab\tmp\passk\*.json; summary table below.
+
+## Headline (honest)
+- **29/29 envs solved by Opus 5.5 with tools; 145/145 attempts correct** (after 3 infra-retries that returned empty,
+  re-run clean). Every model answer was byte-exact equal to our gold (checked separately) — independent frontier
+  confirmation of unique-solvability AND of our oracles.
+- Solve cost: median 44 s, mean 47 s (range 23-121 s), mean 8.4 tool turns. Cross-arch envs (MIPS/AArch64/RISC-V/
+  ARM32/PPC64LE) were NOT harder for Opus than x86-64 once qemu+objdump were available.
+- Implication we state plainly to buyers: the public flag-hunt tier is **warm-up level for a frontier agent**
+  (all-pass groups give ~zero GRPO gradient at that capability level). Our answer: (a) use them as the free demo/
+  calibration tier, (b) sell a **private hard tier** targeting mid-band pass rates for the buyer's own model,
+  (c) offer per-model pass-rate calibration (we run it on our harness before you buy).
+- Not claimed: SOTA comparisons, open-weight numbers yet (next step), pass@N on the private tier (by definition
+  unseen by any model).
+
+## Method notes / gotchas found while building this
+- WSL re-parses argv from Windows (shell metacharacters in answers like `<`, `;`, `&`, backtick mangled) -> grading
+  fixed by generating a tiny POSIX shell script per run (quoting handled by bash file parser). Earlier "0/5" rows for
+  ag-22/24/25/27/28 were OUR harness bug, not model failures — caught and fixed before publishing. (GOTCHAS updated.)
+- Opus `[cyber]` safeguard false-flagged a few prompts; all 145 successful runs used one fresh session each; 3 empty
+  outputs retried.
+
+## Per-env table
+| env | arch | pass@5 | ran_ok any | mean wall s | statuses |
+|---|---|---|---|---|---|
+| brv-ag-01 | x86-64 | 5/5 | YES | 42 | {'OK': 5} |
+| brv-ag-02 | x86-64 | 5/5 | YES | 30 | {'OK': 5} |
+| brv-ag-03 | x86-64 | 5/5 | YES | 37 | {'OK': 5} |
+| brv-ag-04 | x86-64 | 5/5 | YES | 35 | {'OK': 5} |
+| brv-ag-05 | x86-64 | 5/5 | YES | 35 | {'OK': 5} |
+| brv-ag-06 | x86-64 | 5/5 | YES | 37 | {'OK': 5} |
+| brv-ag-07 | x86-64 | 5/5 | YES | 42 | {'OK': 5} |
+| brv-ag-08 | x86-64 | 5/5 | YES | 31 | {'OK': 5} |
+| brv-ag-09 | x86-64 | 5/5 | YES | 43 | {'OK': 5} |
+| brv-ag-10 | x86-64 | 5/5 | YES | 46 | {'OK': 5} |
+| brv-ag-11 | x86-64 | 5/5 | YES | 43 | {'OK': 5} |
+| brv-ag-12 | x86-64 | 5/5 | YES | 45 | {'OK': 5} |
+| brv-ag-13 | x86-64 | 5/5 | YES | 44 | {'OK': 5} |
+| brv-ag-14 | x86-64 | 5/5 | YES | 31 | {'OK': 5} |
+| brv-ag-15 | x86-64 | 5/5 | YES | 51 | {'OK': 5} |
+| brv-ag-16 | x86-64 | 5/5 | YES | 36 | {'OK': 5} |
+| brv-ag-17 | x86-64 | 5/5 | YES | 42 | {'OK': 5} |
+| brv-ag-18 | x86-64 | 5/5 | YES | 46 | {'OK': 5} |
+| brv-ag-19 | x86-64 | 5/5 | YES | 28 | {'OK': 5} |
+| brv-ag-20 | x86-64 | 5/5 | YES | 70 | {'OK': 5} |
+| brv-ag-21 | x86-64 | 5/5 | YES | 54 | {'OK': 5} |
+| brv-ag-22 | mips-eb | 5/5 | YES | 70 | {'OK': 5} |
+| brv-ag-23 | x86-64-pe | 5/5 | YES | 57 | {'OK': 5} |
+| brv-ag-24 | aarch64 | 5/5 | YES | 66 | {'OK': 5} |
+| brv-ag-25 | riscv64 | 5/5 | YES | 78 | {'OK': 5} |
+| brv-ag-26 | x86-64 | 5/5 | YES | 44 | {'OK': 5} |
+| brv-ag-27 | arm-eabi | 5/5 | YES | 67 | {'OK': 5} |
+| brv-ag-28 | ppc64le | 5/5 | YES | 59 | {'OK': 5} |
+| brv-ag-29 | x86-64-pe | 5/5 | YES | 64 | {'OK': 5} |